Cross-border return reason tracing method and device, electronic equipment and storage medium

Abstract

This application relates to the field of tracing the causes of cross-border returns, disclosing a method, apparatus, electronic device, and storage medium for tracing the causes of cross-border returns. The method includes: deeply fusing multilingual text comments and image evidence using joint semantic vector technology; dynamically associating front-end return data with back-end supply chain data through root cause mapping to form a preset root cause map, thereby tracing the causes of returns based on return comment information. The beneficial effects of this invention are: breaking down data silos, accurately identifying deep causal relationships, significantly reducing the root cause misjudgment rate, shortening the root cause location time for cross-border returns, and improving the accuracy of supply chain responsibility allocation, thus achieving precise tracing of the causes of cross-border returns.

Description

Technical Field

[0001] This invention relates to the field of tracing the reasons for cross-border returns, and in particular to a method, apparatus, electronic device, and storage medium for tracing the reasons for cross-border returns. Background Technology

[0002] In the context of globalized e-commerce, cross-border return management faces severe challenges due to the complex supply chain processes involved. Traditional return analysis tools typically employ isolated data processing models: on the one hand, they rely on manual or basic semantic tools to analyze text reviews from different markets; on the other hand, the analysis of user-uploaded images and video evidence of damaged packaging is conducted independently, making it difficult to effectively corroborate textual information. More importantly, this feedback information from the front end is disconnected from the back-end supply chain data, forming data silos. This fragmented approach leads to a one-sided perspective in root cause analysis and a high rate of misjudgment. For example, reviews and images alone might attribute a problem to "defective packaging materials," when the root cause actually lies in the rough handling during the logistics sorting process, but this crucial information is isolated in the logistics system logs and not linked. Summary of the Invention

[0003] Based on this, it is necessary to address the existing problem of tracing the reasons for cross-border returns by proposing a method, device, electronic equipment, and storage medium for tracing the reasons for cross-border returns.

[0004] A method for tracing the reasons for cross-border returns, the method comprising:

[0005] Obtain the return review information of the target product that needs to be traced;

[0006] The target return review information is input into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0007] Before the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0008] Obtain multiple sets of reference return reviews and tags corresponding to each set of reference return reviews; wherein, the tags represent the reference causal relationships between the corresponding reference return reviews.

[0009] Extract the image and text data for each of the aforementioned reference return review information;

[0010] Generate a corresponding joint semantic vector based on each of the aforementioned graphic and textual data;

[0011] The preset root cause graph is generated based on each of the joint semantic vectors and their corresponding labels.

[0012] Furthermore, the step of generating a corresponding joint semantic vector based on each of the image and text data includes:

[0013] For the text content in the reference return review information, a preset large model is used to extract its text semantic features, and for the image content in the reference return review information, a computer vision model is used to identify the visual concepts it contains.

[0014] The text semantic features are vectorized to obtain text vectors, and the visual concepts are vectorized to obtain visual vectors;

[0015] According to the preset weighting rules, the text vector and the visual vector are merged to generate the joint semantic vector.

[0016] Furthermore, after the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0017] Collect data on the actual effects of the improvement strategies implemented based on the aforementioned causal relationships;

[0018] Based on the actual performance data, the model parameters of the preset large model and the computer vision model are fine-tuned using LoRA technology to obtain the updated preset large model and computer vision model.

[0019] Based on the updated preset large model and computer vision model, a joint semantic vector is regenerated;

[0020] The weight assignments in the preset root cause graph are updated based on the regenerated joint semantic vector.

[0021] Furthermore, after the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0022] Obtain multiple regulatory clauses related to the target return review information to form a provisional regulatory database;

[0023] A graded improvement scheme is generated based on the temporary regulatory database and the causal relationship;

[0024] The graded improvement scheme is sent to the first designated terminal.

[0025] Further, the step of generating the preset root cause map based on each of the joint semantic vectors and the corresponding labels includes:

[0026] Extract production batch data and logistics sensor data from the reference causal relationship;

[0027] The joint semantic vector, the production batch data, and the logistics sensing data are associated to construct a multidimensional association matrix;

[0028] Using entities in the joint semantic vector, production batch data, and logistics sensor data as nodes, and causal relationships between entities as edges, an initial graph is constructed.

[0029] The multidimensional correlation matrix and the initial graph are processed using a graph neural network to quantify the influence weights between nodes and generate the preset root cause graph.

[0030] Furthermore, after the step of using a graph neural network to process the multidimensional correlation matrix and the initial graph to quantify the influence weights between nodes and generate the preset root cause graph, the method further includes:

[0031] Acquire multiple sets of time-series data; wherein, the time-series data is data that occurred in the production or logistics process before the return review information;

[0032] The causal strength of the edges of the preset root cause graph is verified using multiple sets of the aforementioned time-series data.

[0033] If the causal strength verification result is a pass, then the preset root cause map is deemed qualified.

[0034] Furthermore, the step of obtaining multiple sets of reference return review information and the tags corresponding to each set of reference return review information includes:

[0035] Obtain multiple sets of historical return review information to form a historical return review information resource pool;

[0036] Each of the historical return review information is input into a preset root cause classification model to obtain the information entropy value of each of the historical return review information;

[0037] Historical return review information with an information entropy value greater than a preset entropy value will be used as reference return review information;

[0038] Based on the reference return review information, it is sent to a second preset terminal to obtain the corresponding tag.

[0039] A device for tracing the reasons for cross-border returns, the device comprising:

[0040] The target return review information acquisition module is used to acquire target return review information that needs to be traced.

[0041] The causal relationship acquisition module is used to input the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0042] The tag acquisition module is used to acquire multiple sets of reference return review information and the tags corresponding to each set of reference return review information; wherein, the tags represent the reference causal relationship of the corresponding reference return review information;

[0043] The image and text data extraction module is used to extract the image and text data for each of the aforementioned reference return review information;

[0044] A joint semantic vector generation module is used to generate a corresponding joint semantic vector based on each of the graphic and text data.

[0045] The preset root cause graph generation module is used to generate the preset root cause graph based on each of the joint semantic vectors and the corresponding labels.

[0046] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0047] Obtain the return review information of the target product that needs to be traced;

[0048] The target return review information is input into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0049] Before the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0050] Obtain multiple sets of reference return reviews and tags corresponding to each set of reference return reviews; wherein, the tags represent the reference causal relationships between the corresponding reference return reviews.

[0051] Extract the image and text data for each of the aforementioned reference return review information;

[0052] Generate a corresponding joint semantic vector based on each of the aforementioned graphic and textual data;

[0053] The preset root cause graph is generated based on each of the joint semantic vectors and their corresponding labels.

[0054] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0055] Obtain the return review information of the target product that needs to be traced;

[0056] The target return review information is input into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0057] Before the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0058] Obtain multiple sets of reference return reviews and tags corresponding to each set of reference return reviews; wherein, the tags represent the reference causal relationships between the corresponding reference return reviews.

[0059] Extract the image and text data for each of the aforementioned reference return review information;

[0060] Generate a corresponding joint semantic vector based on each of the aforementioned graphic and textual data;

[0061] The preset root cause graph is generated based on each of the joint semantic vectors and their corresponding labels.

[0062] The beneficial effects of this invention are as follows: Firstly, by employing joint semantic vector technology to deeply integrate multilingual text comments with image evidence, the problem of fragmented text and image evidence in traditional analysis is solved, improving the accuracy of multimodal evidence matching. Secondly, by using a natural language processing model to extract clauses related to causality from the regulatory database, a tiered improvement scheme is generated based on the matching degree between clauses and causal relationships. Thirdly, by dynamically linking front-end return data with back-end supply chain data through root cause mapping, data silos are broken down, enabling accurate identification of deep causal relationships, significantly reducing the root cause misjudgment rate, shortening the root cause location time for cross-border returns, and improving the accuracy of supply chain responsibility allocation, thereby achieving precise tracing of the causes of cross-border returns. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] in:

[0065] Figure 1 This is a diagram illustrating the application environment of a method for tracing the reasons for cross-border returns in one embodiment.

[0066] Figure 2 This is a flowchart of a method for tracing the reasons for cross-border returns in one embodiment;

[0067] Figure 3 This is a structural block diagram of a device for tracing the reasons for cross-border returns in one embodiment;

[0068] Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Figure 1 This is a diagram illustrating the application environment for tracing the reasons for cross-border returns in one embodiment. (Refer to...) Figure 1 This method for tracing the causes of cross-border returns is applied to a system for tracing the causes of cross-border returns. This system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to acquire the target return review information to be traced, and the server 120 is used to output the causal relationship corresponding to the target return review information.

[0071] like Figure 2 As shown, in one embodiment, a method for tracing the reasons for cross-border returns is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to terminals. The method for tracing the reasons for cross-border returns specifically includes the following steps:

[0072] S1: Obtain the target returned product review information for traceability;

[0073] S2: Input the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0074] Before step S2, which involves inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0075] S101: Obtain multiple sets of reference return review information, and the tags corresponding to each set of reference return review information; wherein, the tags are the reference causal relationships of the corresponding reference return review information;

[0076] S102: Extract the image and text data of each of the aforementioned reference return review information;

[0077] S103: Generate a corresponding joint semantic vector based on each of the aforementioned graphic and textual data;

[0078] S104: Generate the preset root cause map based on each of the joint semantic vectors and the corresponding labels.

[0079] As described in step S1 above, obtain the target returned product review information to be traced. Review information consists of feedback submitted by users after purchasing goods for various reasons (such as product quality, logistics service, discrepancies between description and actual product). Target returned product review information can come from multiple channels, including online retail platforms, feedback forms, and customer service communication records. The system needs to have multilingual processing capabilities to receive reviews from different markets worldwide. To ensure the accuracy and completeness of the data, user-submitted reviews can also undergo data cleaning and preprocessing to remove invalid information, noise, and duplicate content.

[0080] As described in step S2 above, the target return review information is input into a preset root cause graph to obtain the causal relationships corresponding to the target return review information. A root cause graph is a structured, graphical representation that visualizes different events, factors, and their causal relationships, facilitating rapid understanding and tracing of the source of a problem. When inputting review information, the system analyzes sensitive information, keywords, sentiment, and the context of the event in the review text to accurately locate corresponding nodes and edges in the graph. Through this logical connection, the system can depict the causal chain related to the review in the graph. For example, some negative reviews may stem from root causes such as "damaged packaging" or "delayed delivery." In this way, the system can not only identify the direct causes of user dissatisfaction but also help companies uncover the deeper causes of problems, thereby developing more targeted improvement strategies.

[0081] As described in step S101 above, multiple sets of reference return review information and tags corresponding to each set of reference return review information are obtained; wherein, the tags represent the reference causal relationships of the corresponding reference return review information. The tags of each review represent known causal relationships related to the return, such as "logistics issues," "product quality," and "gap in user expectations." These tags are obtained through historical data analysis or manual annotation to ensure their accuracy and reliability. By constructing multiple reference cases, a rich knowledge base can be established, thereby providing reference for new target return reviews.

[0082] As described in step S101 above, the image and text data of each of the reference return review information is extracted. Specifically, key information can be extracted from the text reviews, such as descriptions of item damage, user sentiment, and any additional image or video evidence (e.g., photos of damaged products). The text data is usually preprocessed using Natural Language Processing (NLP) techniques, including word segmentation, stop word removal, and sentiment analysis, to extract valuable information. For image data, computer vision techniques can be applied to identify and classify objects and damage phenomena in the images. This extraction process ensures that the subsequent joint semantic vector generation has a realistic and accurate foundation. By effectively extracting image and text data, the system can not only obtain more context from multimodal information, but also comprehensively consider the information in images and reviews during root cause analysis, forming a more comprehensive understanding and ultimately improving the accuracy and efficiency of the analysis.

[0083] As described in step S103 above, a corresponding joint semantic vector is generated based on each of the text and image data. Specifically, multimodal learning techniques can be used to combine text and image information, so that each data segment is transformed into a high-dimensional vector representation. Specifically, text data is encoded using a pre-trained language model (such as XLM-RoBERTa) to generate its corresponding text vector. At the same time, image data is extracted using an image recognition model (such as CLIP) and transformed into image vectors. Then, the two vectors are integrated and a fusion mechanism (such as weighted summation or concatenation) is used to generate a joint semantic vector, which can comprehensively reflect the features of text and image information. Multimodal fusion and graph construction solve the technical problems of data heterogeneity and correlation. The joint semantic vector fuses text vectors and visual vectors through a cross-modal attention mechanism, where the weight parameters are dynamically learned based on historical data.

[0084] As described in step S104 above, the preset root cause graph is generated based on each of the joint semantic vectors and their corresponding labels. Utilizing the multimodal information reflected by these vector representations, causal relationship chains are established according to the corresponding labels. Specifically, graph neural network technology can be used to construct a multi-dimensional causal graph by using the joint semantic vectors as node features and forming edges between nodes to express the relationships between different events and factors. For example, if the label points to "logistics delay" as the root cause, the relevant joint semantic vectors will form directed edges with other related events (such as "goods damage"), clearly indicating their influence relationship. This not only enhances the expressive power of the graph but also provides a visual basis for root cause analysis. During the construction of the root cause graph, the system can update and optimize its structure in real time, thereby ensuring its accuracy and adaptability to adapt to future changing market demands and user feedback, ultimately improving the intelligence level and efficiency of the entire cross-border return management process.

[0085] In one embodiment, step S103, which generates a corresponding joint semantic vector based on each of the image and text data, includes:

[0086] S1031: Extract the text semantic features of the text content in the reference return review information using a preset large model, and identify the visual concepts contained in the image content in the reference return review information using a computer vision model.

[0087] S1032: Vectorize the text semantic features to obtain text vectors, and vectorize the visual concepts to obtain visual vectors;

[0088] S1033: According to the preset weighting rules, the text vector and the visual vector are fused to generate the joint semantic vector.

[0089] As described in step S1031 above, the text content of the reference return review information is extracted using a preset large model to extract its textual semantic features, and the image content of the reference return review information is identified using a computer vision model to recognize the visual concepts it contains. Natural Language Processing (NLP) is performed on the text content. The preset large model is a multilingual pre-trained language model (such as XLM-RoBERTa). The preset large language model (such as XLM-RoBERTa) is applied to obtain deep semantic features of the text. This process specifically includes word vector generation, analyzing word usage and phrase construction in the review to identify details such as keywords, themes, and sentiment tendencies, thereby extracting meaningful textual semantic features. The computer vision model is a deep learning-based image recognition model (such as CLIP). For image content, the system uses a computer vision model (such as CLIP or CNN deep learning models) to analyze the image to identify the visual concepts it contains. This specifically involves image feature extraction, including object recognition, scene understanding, and detail segmentation. During this process, the model will preprocess the image, such as resizing and normalizing it, and check for common objects (such as damaged packages, goods, etc.).

[0090] As described in step S1032 above, the text semantic features are vectorized to obtain text vectors, and the visual concepts are vectorized to obtain visual vectors. Through the mapping function provided by a large language model, text features are converted into high-dimensional vector representations. In this process, the system can assign a corresponding vector to each text feature point. Each dimension of the vector represents different semantic information. These vectors not only reflect the semantics of the text but also preserve contextual information, giving the text a unique representation in the semantic space. Next, the visual concepts extracted from the image also need to be vectorized. Specifically, this involves using a computer vision model and feature extraction algorithms (such as convolutional neural networks) to transform visual features into numerical forms, generating visual vectors. These vectors represent salient features in the image in a multi-dimensional space, such as color, shape, texture, and object category.

[0091] As described in step S1033 above, the text vector and the visual vector are fused according to a preset weighting rule to generate the joint semantic vector. The text vector and the visual vector are combined using weighted summation, concatenation, or other fusion techniques. For example, assuming the text vector is Vtext and the visual vector is Vimage, the weighting rule can be set as Vcombined = α × Vtext + β × Vimage, where α and β are parameters for adjusting the weights, and α + β = 1. The weights α and β are optimized and determined on the validation set through cross-validation, with initial values ​​set based on modal importance (e.g., text weight 0.6, image weight 0.4). In this way, the system can intelligently adjust the contribution of each modality feature according to different input samples and their characteristics.

[0092] In one embodiment, after step S2 of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0093] S301: Collect data on the actual effects of the improvement strategies implemented based on the causal relationship;

[0094] S302: Based on the actual effect data, the model parameters of the preset large model and the computer vision model are fine-tuned using LoRA technology to obtain an updated preset large model and computer vision model.

[0095] S303: Regenerate the joint semantic vector based on the updated preset large model and computer vision model;

[0096] S304: Update the weight assignment in the preset root cause graph based on the regenerated joint semantic vector.

[0097] As described in step S301 above, collect data on the actual effectiveness of the improvement strategies implemented based on the causal relationship. Improvement strategies can be implemented targeting the root causes of returns, such as replacing unsuitable packaging materials, adjusting logistics partners, or optimizing adjacent supply chain links. Actual effectiveness data includes, but is not limited to, changes in return rates, customer satisfaction survey results, number of complaints, damage rates, and other key performance indicators related to the improvement measures. The data can be collected periodically or in real-time through subsequent user feedback collection, data analysis dashboards, API interfaces, etc. Continuously collecting actual effectiveness data not only allows for the evaluation of the effectiveness of the implemented strategies but also provides valuable references for addressing similar root cause issues in the future.

[0098] As described in step S302 above, based on the actual performance data, the model parameters of the preset large model and the computer vision model are fine-tuned using LoRA technology to obtain updated preset large model and computer vision model. Low-rank adaptation (LoRA) is an effective model fine-tuning method. It achieves efficient parameter updates by decomposing model parameters into combinations of low-rank matrices, reducing computational overhead and avoiding repeated training of the entire model. Specifically, LoRA technology ensures that fine adjustments are made based on the original model, rather than retraining the entire model, which greatly improves efficiency. In this process, the system analyzes the difference between the actual performance data and the model output, identifies the shortcomings of the model in the decision-making process, and optimizes and adjusts the model parameters to improve its adaptability to new data. Specifically, LoRA technology achieves efficient parameter fine-tuning by injecting a low-rank matrix ΔW into the model weights W. The specific formula is W'=W+ΔW, where ΔW=A·B, and A and B are trainable low-rank matrices.

[0099] As described in step S303 above, the joint semantic vector is regenerated based on the updated preset large model and the computer vision model. Based on the new model, features are extracted and processed from the text and previously extracted image content, respectively. The text data undergoes semantic feature re-extraction using the updated large model, while the image data undergoes visual feature re-identification using the improved computer vision model. This regeneration process is similar to the method for generating the joint semantic vector in the previous steps, but the difference is that this generation comprehensively considers the new model parameters and feedback from actual performance data. By effectively integrating text and visual features, the system can generate a new, more representative joint semantic vector to facilitate subsequent root cause mapping updates.

[0100] As described in step S304 above, the weight allocation in the preset root cause graph is updated based on the regenerated joint semantic vector. The root cause graph represents the causal relationships between different events and factors through the connection of nodes. As the model is continuously optimized, the weights in the root cause graph also need to be adjusted accordingly to ensure its ability to reflect the actual situation. Specifically, the relationship between the generated joint semantic vector and the existing root cause graph is analyzed, and nodes that need to have their weights adjusted are identified by calculating similarity or other metrics. Weight updates may involve increasing the influence of a certain causal relationship or reducing certain relationships that have been proven to be irrelevant, thereby maintaining the dynamism and accuracy of the root cause graph. This process not only improves the expressive power of the graph but also ensures that the system can quickly adapt to constantly changing user feedback and market demands.

[0101] In one embodiment, after step S2 of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0102] S311: Obtain multiple regulatory clauses related to the target return review information to form a provisional regulatory database;

[0103] S312: Generate a graded improvement scheme based on the temporary regulatory database and the causal relationship;

[0104] S313: Send the graded improvement scheme to the first designated terminal.

[0105] As described in step S311 above, multiple regulatory clauses related to the target return review information are obtained to form a temporary regulatory database. Based on the target return review information, multiple relevant regulatory clauses are automatically obtained to form a temporary regulatory database. The sources of these regulations may include various laws and regulations. For example, if the review mentions "damaged packaging," the system will focus on querying regulatory clauses related to packaging.

[0106] As described in step S312 above, a tiered improvement plan is generated based on the temporary regulatory database and the causal relationship. The identified regulatory clauses are categorized and prioritized to determine which clauses are most relevant and urgently need to be followed, thereby developing targeted improvement measures. The generated tiered improvement plan typically includes multiple levels of measures. For example, a level 1 improvement plan might involve replacing packaging materials to address packaging damage, while level 2 measures might include negotiations with logistics partners or suggestions for improving transportation processes. The system automatically combines various improvement suggestions into a comprehensive plan, ensuring gradual implementation. This allows companies to select the most effective improvement steps based on their actual situation. The tiered improvement plan generates level 1 (urgent) and level 2 (long-term) improvement measures based on the confidence level of the causal relationship and the severity score of the regulatory violation, using a decision tree model.

[0107] As described in step S313 above, the tiered improvement plan is sent to the first designated terminal. The first designated terminal can be the enterprise's management system, warehouse management system, customer service platform, or it can be sent directly to the mobile device of the relevant responsible person. The sending process needs to ensure the completeness and clarity of the plan content. Therefore, the system should use a concise and clear format and sufficient contextual explanation, and send the specific improvement measures, implementation steps, and relevant legal provisions along with the plan.

[0108] In one embodiment, step S104, which generates the preset root cause map based on each of the joint semantic vectors and the corresponding labels, includes:

[0109] S1041: Extract the production batch data and logistics sensor data from the reference causal relationship;

[0110] S1042: Associate the joint semantic vector, the production batch data, and the logistics sensing data to construct a multidimensional association matrix;

[0111] S1043: Using the entities in the joint semantic vector, production batch data and logistics sensing data as nodes, and the causal relationships between entities as edges, construct an initial graph;

[0112] S1044: Use a graph neural network to process the multidimensional correlation matrix and the initial graph to quantify the influence weights between nodes and generate the preset root cause graph.

[0113] As described in step S1041 above, production batch data and logistics sensor data are extracted from the reference causal relationship. Production batch data typically includes information on raw materials used to produce a specific product, production date, batch number, and usage status of production equipment. This data provides a basis for analyzing quality issues. Logistics sensor data includes various information monitored in real time during transportation, such as transportation temperature, humidity, vibration, impact force, and operational intensity during sorting. This data can reveal potential problems throughout the supply chain.

[0114] As described in step S1042 above, the joint semantic vector, the production batch data, and the logistics sensing data are associated to construct a multidimensional association matrix. The multidimensional association matrix is ​​a high-level feature representation designed to capture the relationships between different data elements. Each data type is standardized to ensure they can interact and be compared within the same numerical range. Subsequently, the system uses appropriate algorithms (such as distance metrics, similarity calculations, etc.) to calculate scores for the association between the joint semantic vector, production batch data, and logistics sensing data. In this process, the system does not simply create one-dimensional associations but expresses the multidimensional relationships between these data through a data structure, allowing each row and column to represent different features and dimensions.

[0115] As described in step S1043 above, an initial graph is constructed using entities from the joint semantic vector, production batch data, and logistics sensor data as nodes, and causal relationships between entities as edges. The nodes of the initial graph include product batches, logistics nodes, and review entities. Edges are initially established using a causal association algorithm (such as the Pearson correlation coefficient). Various entities are identified, such as products mentioned in the target return reviews, batch numbers, names of logistics nodes, and related visual features. Each entity is considered a node in the graph, possessing unique attributes. Then, based on a multidimensional association matrix, the causal relationships between these entities are identified and used as edges to connect related nodes. For example, a product batch might have a high return rate due to "damaged packaging." The system would then establish corresponding edges between these nodes to represent their causal relationships. In this way, the graph can effectively demonstrate the mutual influence between different factors, forming a causal relationship chain.

[0116] As described in step S1044 above, the multidimensional association matrix and the initial graph are processed using a graph neural network to quantify the influence weights between nodes and generate the preset root cause graph. The multidimensional association matrix is ​​used as the node feature input to the graph neural network. The initial graph structure is constructed based on entity relationships. GNN is a deep learning technique widely used in graph data analysis. By propagating information between nodes, the system can capture complex nonlinear relationships and graph structure information. The core advantage of GNN lies in its ability to consider the features of each node and the information of its neighboring nodes, thereby calculating a meaningful influence weight for each node. Specifically, the information in the multidimensional association matrix is ​​used as input, and the initial graph is used as the structural connection. During each convolutional layer, the GNN updates the feature representation of each node. That is, during each message passing layer, the GNN updates the current node representation by aggregating the features of neighboring nodes, and then extracts node features by fusing the features of neighboring nodes. These features are converted into influence weights, representing the importance and degree of influence of a specific node in the overall causal chain.

[0117] In one embodiment, after step S1044, which involves processing the multidimensional correlation matrix and the initial graph using a graph neural network to quantify the influence weights between nodes and generate the preset root cause graph, the method further includes:

[0118] S10451: Acquire multiple sets of time-series data; wherein, the time-series data is data that occurred in the production or logistics process before the return review information;

[0119] S10452: Use multiple sets of the aforementioned time-series data to perform causal strength verification on the edges of the preset root cause graph;

[0120] S10453: If the result of the causal strength verification is that the verification is passed, then the preset root cause map is deemed qualified.

[0121] As described in step S10451 above, multiple sets of time-series data are acquired. The time-series data contains all information related to returns during the production or logistics process, and these data occurred before the return comments were generated. Data can be extracted from multiple data sources, such as sensor data from production line monitoring equipment, delivery information from the logistics system, and inbound and outbound records from the warehouse management system. Time-series data types include changes in production batches, environmental monitoring data (such as temperature and humidity), and various sensor data during transportation (such as speed, vibration, and impact).

[0122] As described in step S10452 above, the causal strength of the edges in the preset root cause graph is verified using multiple sets of time-series data. Causal strength verification typically employs statistical analysis and machine learning methods to ensure consistency between the model output and the actual collected data. Specifically, the system uses model testing, regression analysis, or other algorithms suitable for handling causal relationships to conduct a detailed analysis of each causal edge. During this process, the causal strength of the edge is measured to determine whether there is a significant correlation between previously occurring event data (such as production conditions or logistics processes) and the target return review data. For example, if an edge represents "excessive vibration of sorting equipment leading to product damage," the time-series data needs to show that vibration data did exceed the standard within a certain time period, and that corresponding return records were generated within that period. Through this verification process, the system can further confirm which causal relationships are valid and which may only be superficial phenomena, thereby ensuring the scientific validity and reliability of the root cause graph.

[0123] As described in step S10453 above, if the causal strength verification passes, the preset root cause graph is deemed qualified. If the verification result shows that the strength of each causal relationship edge meets expectations and there is significant supporting data, this indicates that the formed root cause graph is reasonable and effective in terms of structure and information expression. At this time, the system will determine the root cause graph as "qualified." This determination process is a key step in the entire root cause analysis, which is related to whether the formulation and implementation of subsequent improvement strategies are scientific and effective. If the graph is determined to be qualified, the system will retain and use the graph to guide future decision-making and strategy generation, ensuring that the existing root cause analysis framework can be quickly applied when facing new return events. However, if the causal strength verification fails, the system can make adjustments in reverse, re-examining the relevant formulas, algorithms, or data acquisition process to ensure that the information and structure in the graph are accurate enough and have practical value.

[0124] In one embodiment, step S101, which involves obtaining multiple sets of reference return review information and the tag corresponding to each set of reference return review information, includes:

[0125] S1011: Obtain multiple sets of historical return review information to form a historical return review information resource pool;

[0126] S1012: Input each of the historical return review information into a preset root cause classification model to obtain the information entropy value of each of the historical return review information;

[0127] S1013: Use historical return review information with an information entropy value greater than a preset entropy value as reference return review information;

[0128] S1014: Based on the reference return comment information, send it to the second preset terminal to obtain the corresponding tag.

[0129] As described in step S1011 above, multiple sets of historical return review information are acquired to form a historical return review information resource pool. Historical return review information can come from various channels, such as online retail platforms, consumer feedback forms, customer service records, and social media. During the data acquisition process, the system can also clean and standardize the acquired data to remove invalid information, duplicate data, and noise, ensuring the high quality of the final resource pool.

[0130] As described in step S1012 above, each historical return review is input into a preset root cause classification model to obtain the information entropy value of each historical return review. The root cause classification model, after training, can understand and classify different return reviews, identifying key factors. Information entropy, in information theory, is a measure used to quantify uncertainty or information content, reflecting the complexity and importance of data. By calculating the output probability, the information entropy value of each review can be obtained. A high information entropy value means that the review contains rich information and high uncertainty, helping the system identify reviews with greater potential value as reference data. The information entropy value... ,in For the root cause classification model, categories, The output probability.

[0131] As described in step S1013 above, historical return reviews with information entropy values ​​greater than a preset entropy value are used as reference return review information. The information entropy values ​​of all historical return reviews are compared with the preset entropy value to identify those reviews with information entropy values ​​greater than the preset entropy value. These reviews are generally considered to be rich in content, have a large amount of information, and have high uncertainty in understanding. They can provide the system with more insights and analytical value. The preset entropy value can be set to 0.7. High information entropy reviews are selected as references because these reviews often better reflect real consumer feedback and potential fundamental problems, such as complex reasons for returns, customer emotions, or specific expectations and disappointments with the product. Through this virtual screening, the system can efficiently extract the most valuable reviews for research and avoid diluting the analysis effect with a large number of low-value, monotonous reviews.

[0132] As described in step S1014 above, the reference return review information is sent to a second preset terminal to obtain corresponding tags. The second preset terminal can be a human review team responsible for data annotation, a knowledge base, or another data processing system. Its purpose is to obtain tags corresponding to these review information. These tags generally contain specific information related to the content and sentiment of each review, possibly including the root cause, problem category, customer satisfaction, etc. During this process, to ensure that the sent data is specific and easy to understand, necessary contextual information can also be attached to assist the terminal in quickly and accurately annotating. In this way, the system can efficiently concentrate human resources, allowing them to focus on processing high-value data, while ensuring the quality and consistency of the annotation.

[0133] Reference Figure 3 The present invention also provides a device for tracing the reasons for cross-border returns, the device comprising:

[0134] The target return review information acquisition module 902 is used to acquire target return review information that needs to be traced.

[0135] The causal relationship acquisition module 904 is used to input the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0136] The tag acquisition module 906 is used to acquire multiple sets of reference return review information and tags corresponding to each set of reference return review information; wherein, the tags are reference causal relationships for the corresponding reference return review information;

[0137] The image and text data extraction module 908 is used to extract the image and text data of each of the reference return review information;

[0138] The joint semantic vector generation module 910 is used to generate a corresponding joint semantic vector based on each of the graphic data;

[0139] The preset root cause graph generation module 912 is used to generate the preset root cause graph based on each of the joint semantic vectors and the corresponding labels.

[0140] In one embodiment, the joint semantic vector generation module 910 includes:

[0141] The visual concept recognition submodule is used to extract the text semantic features of the text content in the reference return review information using a preset large model, and to recognize the visual concepts contained in the image content in the reference return review information using a computer vision model.

[0142] The visual vector acquisition submodule is used to vectorize the text semantic features to obtain text vectors, and to vectorize the visual concepts to obtain visual vectors.

[0143] The joint semantic vector generation submodule is used to generate the joint semantic vector by fusing the text vector and the visual vector according to a preset weighting rule.

[0144] In one embodiment, the device for tracing the reasons for cross-border returns further includes:

[0145] The actual effect data collection module is used to collect actual effect data of the improvement strategies implemented based on the causal relationship;

[0146] The fine-tuning module is used to fine-tune the model parameters of the preset large model and the computer vision model based on the actual effect data, using LoRA technology, so as to obtain the updated preset large model and computer vision model.

[0147] A joint semantic vector regeneration module is used to regenerate joint semantic vectors based on the updated preset large model and computer vision model.

[0148] An update module is used to update the weight allocation in the preset root cause graph based on the regenerated joint semantic vector.

[0149] In one embodiment, the device for tracing the reasons for cross-border returns further includes:

[0150] A temporary regulatory database formation module is used to obtain multiple regulatory clauses related to the target return review information to form a temporary regulatory database;

[0151] A graded improvement scheme generation module is used to generate graded improvement schemes based on the temporary regulatory database and the causal relationship.

[0152] The sending module is used to send the graded improvement scheme to the first designated terminal.

[0153] In one embodiment, the root cause mapping generation module 912 includes:

[0154] The extraction submodule is used to extract production batch data and logistics sensor data from the reference causal relationship;

[0155] The association submodule is used to associate the joint semantic vector, the production batch data and the logistics sensing data to construct a multi-dimensional association matrix;

[0156] A submodule is constructed to build an initial graph using entities in the joint semantic vector, production batch data and logistics sensor data as nodes and causal relationships between entities as edges.

[0157] The quantization submodule is used to process the multidimensional correlation matrix and the initial graph using a graph neural network to quantify the influence weights between nodes and generate the preset root cause graph.

[0158] In one embodiment, the root cause mapping generation module 912 further includes:

[0159] The time-series data acquisition submodule is used to acquire multiple sets of time-series data; wherein, the time-series data is data that occurred in the production or logistics process before the return review information;

[0160] The verification submodule is used to verify the causal strength of the edges of the preset root cause graph using multiple sets of time-series data.

[0161] The determination submodule is used to determine that the preset root cause map is qualified if the result of the causal strength verification is qualified.

[0162] In one embodiment, the tag acquisition module 906 includes:

[0163] The historical return review information acquisition submodule is used to acquire multiple sets of historical return review information to form a historical return review information resource pool;

[0164] The information entropy value acquisition submodule is used to input each of the historical return review information into a preset root cause classification model to obtain the information entropy value of each of the historical return review information;

[0165] The reference return review information acquisition submodule is used to use historical return review information with an information entropy value greater than a preset entropy value as reference return review information;

[0166] The tag acquisition submodule is used to send the reference return review information to a second preset terminal to obtain the corresponding tag.

[0167] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a method for tracing the causes of cross-border returns. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement a method for tracing the causes of cross-border returns. Those skilled in the art will understand that... Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0168] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0169] Obtain the return review information of the target product that needs to be traced;

[0170] The target return review information is input into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0171] Before the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0172] Obtain multiple sets of reference return reviews and tags corresponding to each set of reference return reviews; wherein, the tags represent the reference causal relationships between the corresponding reference return reviews.

[0173] Extract the image and text data for each of the aforementioned reference return review information;

[0174] Generate a corresponding joint semantic vector based on each of the aforementioned graphic and textual data;

[0175] The preset root cause graph is generated based on each of the joint semantic vectors and their corresponding labels.

[0176] By employing joint semantic vector technology to deeply integrate multilingual text comments with image evidence, the problem of fragmented text and image evidence in traditional analysis is solved, thereby improving the accuracy of multimodal evidence matching. Secondly, by using root cause mapping to dynamically link front-end return data with back-end supply chain data, data silos are broken down, enabling accurate identification of deep causal relationships, significantly reducing the root cause misjudgment rate, shortening the root cause location time for cross-border returns, and improving the accuracy of supply chain responsibility allocation, thus achieving precise tracing of the causes of cross-border returns.

[0177] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps:

[0178] Obtain the return review information of the target product that needs to be traced;

[0179] The target return review information is input into a preset root cause graph to obtain the causal relationship corresponding to the target return review information;

[0180] Before the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes:

[0181] Obtain multiple sets of reference return reviews and tags corresponding to each set of reference return reviews; wherein, the tags represent the reference causal relationships between the corresponding reference return reviews.

[0182] Extract the image and text data for each of the aforementioned reference return review information;

[0183] Generate a corresponding joint semantic vector based on each of the aforementioned graphic and textual data;

[0184] The preset root cause graph is generated based on each of the joint semantic vectors and their corresponding labels.

[0185] By employing joint semantic vector technology to deeply integrate multilingual text comments with image evidence, the problem of fragmented text and image evidence in traditional analysis is solved, thereby improving the accuracy of multimodal evidence matching. Secondly, by using root cause mapping to dynamically link front-end return data with back-end supply chain data, data silos are broken down, enabling accurate identification of deep causal relationships, significantly reducing the root cause misjudgment rate, shortening the root cause location time for cross-border returns, and improving the accuracy of supply chain responsibility allocation, thus achieving precise tracing of the causes of cross-border returns.

[0186] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0187] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0188] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for tracing the reasons for cross-border returns, characterized in that, The method includes: Obtain the return review information of the target product that needs to be traced; The target return review information is input into a preset root cause graph to obtain the causal relationship corresponding to the target return review information; Obtain multiple regulatory clauses related to the target return review information to form a provisional regulatory database; A graded improvement scheme is generated based on the temporary regulatory database and the causal relationship; Send the graded improvement scheme to the first designated terminal; Before the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes: Obtain multiple sets of reference return reviews and tags corresponding to each set of reference return reviews; wherein, the tags represent the reference causal relationships between the corresponding reference return reviews. Extract the image and text data for each of the aforementioned reference return review information; Generate a corresponding joint semantic vector based on each of the aforementioned graphic and textual data; The preset root cause graph is generated based on each of the joint semantic vectors and the corresponding labels; The step of generating the preset root cause graph based on each of the joint semantic vectors and the corresponding labels includes: Extract production batch data and logistics sensor data from the reference causal relationship; The joint semantic vector, the production batch data, and the logistics sensing data are associated to construct a multidimensional association matrix; Using entities in the joint semantic vector, production batch data, and logistics sensor data as nodes, and causal relationships between entities as edges, an initial graph is constructed. The multidimensional correlation matrix and the initial graph are processed using a graph neural network to quantify the influence weights between nodes and generate the preset root cause graph.

2. The method for tracing the reasons for cross-border returns according to claim 1, characterized in that, The step of generating a corresponding joint semantic vector based on each of the image and text data includes: For the text content in the reference return review information, a preset large model is used to extract its text semantic features, and for the image content in the reference return review information, a computer vision model is used to identify the visual concepts it contains. The text semantic features are vectorized to obtain text vectors, and the visual concepts are vectorized to obtain visual vectors; According to the preset weighting rules, the text vector and the visual vector are merged to generate the joint semantic vector.

3. The method for tracing the reasons for cross-border returns according to claim 2, characterized in that, After the step of inputting the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information, the method further includes: Collect data on the actual effects of the improvement strategies implemented based on the aforementioned causal relationships; Based on the actual performance data, the model parameters of the preset large model and the computer vision model are fine-tuned using LoRA technology to obtain the updated preset large model and computer vision model. Based on the updated preset large model and computer vision model, a joint semantic vector is regenerated; The weight assignments in the preset root cause graph are updated based on the regenerated joint semantic vector.

4. The method for tracing the reasons for cross-border returns according to claim 1, characterized in that, After the step of using a graph neural network to process the multidimensional correlation matrix and the initial graph to quantify the influence weights between nodes and generate the preset root cause graph, the method further includes: Acquire multiple sets of time-series data; wherein, the time-series data is data that occurred in the production or logistics process before the return review information; The causal strength of the edges of the preset root cause graph is verified using multiple sets of the aforementioned time-series data. If the causal strength verification result is a pass, then the preset root cause map is deemed qualified.

5. The method for tracing the reasons for cross-border returns according to claim 1, characterized in that, The steps of obtaining multiple sets of reference return reviews and the corresponding tags for each set of reference return reviews include: Obtain multiple sets of historical return review information to form a historical return review information resource pool; Each of the historical return review information is input into a preset root cause classification model to obtain the information entropy value of each of the historical return review information; Historical return review information with an information entropy value greater than a preset entropy value will be used as reference return review information; Based on the reference return review information, it is sent to a second preset terminal to obtain the corresponding tag.

6. A device for tracing the reasons for cross-border returns, characterized in that, The device includes: The target return review information acquisition module is used to acquire target return review information that needs to be traced. The causal relationship acquisition module is used to input the target return review information into a preset root cause graph to obtain the causal relationship corresponding to the target return review information; A temporary regulatory database formation module is used to obtain multiple regulatory clauses related to the target return review information to form a temporary regulatory database; A graded improvement scheme generation module is used to generate graded improvement schemes based on the temporary regulatory database and the causal relationship. The sending module is used to send the graded improvement scheme to the first designated terminal; The tag acquisition module is used to acquire multiple sets of reference return review information and the tags corresponding to each set of reference return review information; wherein, the tags represent the reference causal relationship of the corresponding reference return review information; The image and text data extraction module is used to extract the image and text data for each of the aforementioned reference return review information; A joint semantic vector generation module is used to generate a corresponding joint semantic vector based on each of the graphic and text data. A preset root cause graph generation module is used to generate the preset root cause graph based on each of the joint semantic vectors and the corresponding labels. The preset root cause map generation module includes: The extraction submodule is used to extract production batch data and logistics sensor data from the reference causal relationship; The association submodule is used to associate the joint semantic vector, the production batch data and the logistics sensing data to construct a multi-dimensional association matrix; A submodule is constructed to build an initial graph using entities in the joint semantic vector, production batch data and logistics sensor data as nodes and causal relationships between entities as edges. The quantization submodule is used to process the multidimensional correlation matrix and the initial graph using a graph neural network to quantify the influence weights between nodes and generate the preset root cause graph.

7. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the method for tracing the cause of cross-border returns as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method for tracing the cause of cross-border returns as described in any one of claims 1 to 5.

Images